[mps/inductor] Adjust lowering to not emit comments

[dcci]

🐛 Describe the bug

Edit
The problem weren't VLAs, rather, the fact that // is a comment in metal and an operation in python.

** Original report **
There are currently a fair amount of inductor tests that are failing because the lowering emits VLAs, which aren't supported by the Metal shading language (according to https://developer.apple.com/metal/Metal-Shading-Language-Specification.pdf -- only fixed sized arrays are supported).

We will need to massage the codegen a little bit to not emit VLAs, or find some alternative solution. Filing this issue so that I don't forget.

cc: @malfet

cc @kulinseth @albanD @malfet @DenisVieriu97 @jhavukainen @chauhang @penguinwu

[dcci]

Example:

program_source:19:9: error: expected ')'
        auto tmp10 = tmp8 ? tmp9 : 0.0;
        ^
program_source:18:29: note: to match this '('
        auto tmp9 = in_ptr0[((x0 - 2)//3)];
                            ^ 
program_source:19:39: error: expected ']'
        auto tmp10 = tmp8 ? tmp9 : 0.0;
                                      ^
program_source:18:28: note: to match this '[' 
        auto tmp9 = in_ptr0[((x0 - 2)//3)];
                           ^
program_source:21:29: error: use of undeclared identifier 'tmp10'
        auto tmp12 = tmp8 ? tmp10 : tmp11;
                            ^ 

from: GPUTests.test_slice_scatter4_mps.

The parser error isn't particularly informative but I believe what it's trying to say is that x0 / 2 isn't known at compile time so that makes tmp9 a VLA.

[malfet]

I believe this is different, but I've though I've fixed it

[malfet]

Q to the oncall: should I mark those as PT2 issues or not?

[malfet]

@dcci example you've posted are not because of VLA, but because // is a floordiv in Python, but a comment in C/C++/Metal

        auto tmp9 = in_ptr0[((x0 - 2)//3)];
        auto tmp10 = tmp8 ? tmp9 : 0.0;

will indeed result in (see syntax highlighting, Metal thinks that everything after // is a comment

program_source:18:29: note: to match this '('
        auto tmp9 = in_ptr0[((x0 - 2)//3)];
                            ^ 
program_source:19:39: error: expected ']'
        auto tmp10 = tmp8 ? tmp9 : 0.0;

Fix is coming

[dcci]

hmm, makes sense, but I believe I did find a case where this failed because of VLA. Before closing this issue, let me try digging into that and see if I can find it tomorrow.